As solar energy becomes increasingly critical to thepower grid, the ability to predict short-term fluctuations ingeneration is essential for maintaining stability and efficiency.While many studies propose advanced machine learning models,few address the engineering challenges of deploying these modelsinto live, production-oriented prototypes. This paper presentsan end-to-end, cloud-deployed AI system for next-hour solarpower forecasting, constructed utilizing real-world generationdata from a plant in Bhopal, India, synthesized with satellitederivedweather data from NASA POWER. By engineeringtime-cyclic and lag features to capture temporal dependencies,we trained a Random Forest Regressor that achieves a MeanAbsolute Error (MAE) of approximately 1406W and a RootMean Squared Error (RMSE) of 2768W. Beyond the modelingphase, this work focuses on the system architecture required tomake these predictions accessible and reliable. We implementa lightweight MLOps pipeline to operationalize the model as acloud-deployed service, demonstrating deployability constraintson low-cost infrastructure. The proposed approach reducesRMSE by approximately 35% relative to a persistence baseline,with performance stability verified across multiple random seeds(n = 5).
Varsh Vishwakarma (Sat,) studied this question.